Abstract: Machine Learning algorithms can be described as read-only methods, because they generate a representation and/or a model of the data, without changing it. Given that the quality of the model learned is strongly dependent on the alignment between the data and the bias of the algorithm, this means an essential part of the process of developing ML models is data preparation. However, data preparation methods are essentially independent of the learning algorithm used, which makes it hard to ensure that the changes performed to the data are aligned with the algorithm that will later be used to learn a model. To address this fundamental disalignment, we have recently proposed a new paradigm for ML, read-write ML (rwML). rwML algorithms extend existing ones with data preparation abilities and are, thus, able not only to learn models but to change the training data in order to obtain better ones. We have proposed a few rwML algorithms with promising results. In this talk, I will discuss them as well as some ideas for the future of rwML.
Speaker: Professor Carlos Soares, FEUP
Location and Date: IEETA auditorium, 12 April 2024, 14h
